ABSTRACT: Today’s real-world physical systems such as smart buildings and factories,  as well as many advanced machineries such as cranes and mining equipment, are becoming large and complex data-intensive systems. Constant monitoring and analysis of the data streams generated by the multitude of interconnected sensors and actuators in these systems can be useful for intelligent fault diagnosis or for detecting anomalies due to possible cyber intrusions, using advanced machine learning approaches such as deep learning.  However, detecting anomalous patterns in multivariate time series, which is the fundamental data type for understanding dynamics of the underlying processes in these real-world systems, is a challenging problem. The deep learning models need to address various issues such as accounting for the complicated temporal and spatial dynamics of the sensors present in these complex systems, and addressing possible variations in the working conditions between training and testing time, or across different machineries and cyber-physical systems, through domain adaption. We present our ongoing efforts in intelligent fault diagnosis under varying working conditions based on domain adaptive convolutional neural networks, and multivariate anomaly detection with self-learning graph convolutional networks for cyber-physical systems.

 

intelligent fault diagnosis, anomaly detection, cyber physical systems, convolutional neural networks, deep learning